writing-cc

Writing-CC: Automated Document Writing with Iterative Review

Write and refine: $ARGUMENTS

Overview

Use this skill when the user wants a polished, well-researched written document on any topic. The skill does not simply write and hand off. It first understands what makes the document valuable, derives measurable evaluation criteria, writes a grounded first draft, then runs it through an external GPT reviewer using those exact criteria — iterating until the document meets the bar or MAX_ROUNDS is reached.

Three principles dominate:

1. Requirements first, writing second. Extract what the user actually needs before producing a single word of content.
2. Criteria-driven quality. Every evaluation round uses the same rubric derived from the user's requirements — not generic writing advice.
3. Grounded writing. All claims and examples must come from real sources: web search, local materials, or cited papers. Never fabricate examples.

User input (document request)
  -> Phase 0 (Claude): Analyze requirements — purpose, audience, key points, constraints
  -> Phase 1 (Claude): Derive evaluation rubric (5-7 measurable criteria) from requirements
  -> Phase 2 (Claude): Research (web search + local materials) + write first draft
  -> Phase 3 (Codex/GPT): Score draft on rubric, provide structured feedback
  -> Phase 4 (Claude): Parse feedback, revise document
  -> Phase 5 (Codex, same thread): Re-score revised document
  -> Repeat Phase 4-5 until OVERALL SCORE >= SCORE_THRESHOLD or MAX_ROUNDS reached
  -> Phase 6: Save full history to writing-logs/

Constants

• REVIEWER_MODEL = gpt-5.4 — Reviewer model used via Codex MCP.
• MAX_ROUNDS = 5 — Maximum review-revise cycles.
• SCORE_THRESHOLD = 8.5 — Minimum overall score to stop early.
• OUTPUT_DIR = writing-logs/ — Directory for all output files.
• MAX_SEARCH_QUERIES = 10 — Maximum web search queries during research.
• MAX_LOCAL_FILES = 10 — Maximum local files to scan for grounding.